Research On Automatic Segmentation And Classification Of Thyroid Nodule Ultrasound Images

With the development of society,the effect of work stress,diet,daily schedule,radiation has increased significantly.As a consequence,thyroid diseases are becoming more common and the incidence is rising continuously.It has great significant to circle the lesion area and obtain the result of benign or malignant in the thyroid ultrasound images.In this paper,deep learning methods and traditional image processing methods are used on thyroid ultrasound images.Additionally,a semi-supervised learning segmentation algorithm is proposed for the small amount of medical image data to ease the dependence of the algorithm on large data volumes.The following is a detailed description of the work in this article:1.Segmentation of thyroid nodules:Firstly,deep neural network is applied to segment thyroid nodules.The new network is mainly obtained by improving the U-net++network.The accuracy of the segmentation result is DICE accuracy of 68.11%.The second step is to further refine the nodules obtained from deep learning,and use Snake algorithm to evolve the deep learning segmentation results to improve the accuracy of segmentation.In many algorithms,the focus is only on the improvement of deep networks.While improving the structure of deep learning networks,this paper analyzes the shortcomings of the output results of deep learning algorithms,and adds a secondary segmentation algorithm to improve the accuracy.After two divisions,the DICE accuracy of the final result reached 70.65%.2.Classification of benign and malignant thyroid nodules:The methods of traditional image classification and deep learning classification are used to classify benign and malignant nodules.The traditional classification method first extracts the image features,and then sends the features to the classifier for classification.This method achieves 87.58%of F1 score in the test set.Convolutional neural networks are used to classify images in deep learning.By improving the high-resolution Net(HRNet),the F1 score of this algorithm on the test set is increased from 73.34%of the original HRNet to 76.16%.The effect of deep learning classification does not exceed that of traditional classification methods.It is due to the large amount of data required for deep learning,but the amount of the data set used in this experiment is small.3.Semi-supervised learning for segmentation of thyroid nodules:Aiming at the problem that current medical images and labels are difficult to obtain a semi-supervised learning method is proposed.Only a part of the data label and another part of the long and short axis information can be used to obtain the model.Although there are many researches on semi-supervised learning,it is rarely explored in the field of thyroid nodule segmentation.In the experiments,the weak supervised learning methods in other papers were referenced,and its iterative criteria and data usage methods were changed to make it more suitable for thyroid ultrasound data.The DICE accuracy of the final segmentation result using only part of the supervised information reached 66.8%.